Abstract
This paper presents a theory of algorithms designed to optimize highly interactive systems (multi-diminsional, multi-peak, nonlinear functions). Two applications are discussed: one concerns cognitive systems capable of learning and generalization, and one concerns calculations dealing with the "origin of life" from "organic soups". The algorithms are intrinsically parallel--each function argument processed serves as a carrier for information about a tremendous number of regions (hyperplanes) in the function's domain. Each region is automatically ranked according to the estimated average value of the function over that region, and the rankings are compactly stored in the algorithm's data base (M l-tuples store approximately M-2 l/2 rankings). Thus, the algorithm implicitly processes hyperplanes by manipulating function arguments.